Search

KR-20260062964-A - Method for reducing image variability under global and local illumination variations

KR20260062964AKR 20260062964 AKR20260062964 AKR 20260062964AKR-20260062964-A

Abstract

The present invention provides a system, device, and method(s) for processing an image based on color constancy. The method comprises the steps of: acquiring an image in a first color space; converting the image into data in a second color space; converting the data using color adaptation; performing a first normalization on the converted data, wherein the first normalization includes applying a dynamic spatial filtering technique to adjust the converted data based on light intensity; applying a filter set to the normalized data, wherein the filter set is convolutionally processed based on the normalized data regarding the image; performing a second normalization on the filtered data to obtain an illumination estimate of the image regarding the filtered data; and outputting normalized data from the second normalization, wherein the normalized data maintains color constancy based on the illumination estimate and removes illumination from the normalized data.

Inventors

  • 마보우디 하디

Assignees

  • 옵테란 테크놀로지스 리미티드

Dates

Publication Date
20260507
Application Date
20240829
Priority Date
20230901

Claims (20)

  1. A computer-implemented method for image processing that removes illumination from an image based on color constancy, wherein the method comprises: Step of acquiring an image in a first color space; A step of converting the above image into data in a second color space; A step of converting the above data using color adaptation; A step of performing a first normalization on the transformed data, wherein the first normalization includes assigning weights to the transformed data based on the cumulative distribution of light intensity values of the image; A step of applying a filter set to the normalized data, wherein the filter set is convolutionally processed based on the normalized data relating to the image; A step of obtaining an illumination estimate of an image regarding filtered data and performing a second normalization on image data in a second color space by dividing image data in a second color space by said illumination estimate; and A computer-implemented method comprising the step of outputting normalized data from the second normalization, wherein the normalized data maintains color equilibrium based on the light estimation value and the step of removing light from the normalized data.
  2. In paragraph 1, the above method is: A method further comprising the step of converting normalized data from a second normalization into an image in a first color space.
  3. In paragraph 1 or 2, the above method is: The step of remapping the transformed data into a sparse matrix further includes a step in which each position of the first dimension corresponds to a pixel of the image and the second dimension corresponds to a range of possible pixel values. Here, the first normalization is a method performed on the remapped data.
  4. A method according to any one of claims 1 to 3, wherein the second normalization includes correcting data filtered by an illumination vector generated using a pooling function, wherein the filtered data is divisible by an illumination vector.
  5. In paragraph 4, the pooling function includes the following: Here, f (x,y) (.) is the data representation of the normal neural operation of the maximum value for the filtered data, and A method representing the output of a double-opposition filter in a second color space.
  6. In any one of paragraphs 1 to 5, the method is: A step of receiving raw images from one or more cameras; Step of removing gamma correction from the raw image; and A method comprising further a step of converting a raw image into an image in a first color space.
  7. A method according to any one of claims 1 to 6, wherein the first color space is a red-green-blue color space.
  8. A method according to any one of claims 1 to 7, wherein the second color space is a long-medium-short color space.
  9. In any one of claims 1 to 8, the step of converting the data using color adaptation is: Step of estimating the light source of the data using a gray world model; and A method comprising further a step of converting data into a target light source using the above light source.
  10. In paragraph 9, the step of converting the above data into a target light source is: A method comprising the step of applying one or more matrix transformations to a target light source according to one or more cameras for image acquisition.
  11. A method according to any one of claims 1 to 10, wherein the filter set comprises at least one center-periphery structure representing an RF encoding a color contrast.
  12. In paragraph 11, RF is a method comprising red-green contrast, blue-yellow contrast and achromatic contrast.
  13. A method according to any one of claims 1 to 12, wherein the filter set comprises at least two filters arranged in series such that one of the at least two filters receives an input from another filter.
  14. A method according to any one of claims 1 to 13, wherein the filter set represents two layers of a visual system.
  15. In any one of claims 1 to 14, the method is: Step of acquiring an input image; A step of dividing the above input image into multiple regions; A step of analyzing the plurality of regions based on the color information and spatial location of the pixels of each region; A step of selecting a subset of regions affected by a light source from the plurality of regions based on the above analysis; A step of identifying color edges for at least a subset of the above-mentioned region; A step of extracting color information from the above color edge; A step of separating the reflectance component and the illumination component of the input image using the extracted color information above; A step of correcting the illumination of the input image based on the separated reflectance component and illumination component; A method comprising an additional step of outputting a lighting-corrected image.
  16. A method according to claim 15, wherein the acquired input image is a raw image received from one or more cameras, the image in a first color space acquired prior to first normalization, the normalized data from the output of second normalization, or the image in a first color space acquired after second normalization.
  17. In paragraph 15 or 16, the step of analyzing the plurality of regions is: A step of identifying prominent regions from multiple regions using a clustering algorithm; A step of generating a clustered map based on identified prominent regions; and A method comprising the step of analyzing a prominent region to select a subset of the region affected by a light source based on a clustered map.
  18. In paragraph 17, a method wherein the clustered map limits multiple regions to a set number of partitions.
  19. A method according to any of claims 15 to 18, wherein the step of dividing an input image into a plurality of regions further comprises the step of dividing the input image using a k-means clustering algorithm based on the similarity of color or intensity values of pixels of the input image.
  20. In any of claims 15 to 19, the step of identifying color edges for at least a subset of the region further comprises the step of performing edge detection using a Canny edge detection algorithm configured to select a plurality of edges based on threshold intensity.

Description

Method for reducing image variability under global and local illumination variations The present application relates to a system, apparatus, and method(s) for estimating local and global illumination of an image, maintaining color constancy, and minimizing shadow effects of an image by removing illumination through color correction. Color constancy is the ability to perceive the color of an object regardless of changes in the light source. This ability is reported to exist inherently in species such as humans, fish, and bees. In these species, color constancy has evolved to assist object recognition by reducing memory consumption while maintaining the same or superior accuracy and stable discrimination. Without color constancy, object colors under varying lighting conditions would be difficult to trust, thereby degrading the accurate object identification ability of the species. Numerous behavioral and neurobiological studies have sought to elucidate the neural mechanisms underlying color constancy. Two neural mechanisms explaining color constancy have been proposed: chromatic adaptation and the opponent-process theory. Chromatic adaptation occurs at the photoreceptor level by regulating the activity of individual photoreceptors based on the relative light intensity within a local area. High-level neural processing, represented by double-opponent cells in the early visual system (regions V1 and V4), is also believed to contribute to color constancy. Meanwhile, the opponent-process theory suggests that one of the two colors in a color pair inhibits the other. In the fields of computer vision and robotics, a critical requirement, particularly for robust color-based object recognition and tracking, is the ability to record and store reliable color cues that remain invariant to changes in external lighting. Meeting this requirement becomes increasingly difficult when the light source is unknown or when heterogeneous lighting conditions exist simultaneously across a scene. In the field of computer color constancy, various solutions have been proposed regarding robust color coding techniques that correct color-biased images to obtain canonical images under a white light source. In this regard, numerous biologically valid and ad hoc solutions have been proposed for computer color constancy, ranging from simple algorithms based on low- or intermediate-level image feature statistics (e.g., grey-world models, white patches, max-RGB, shades of grey, grey edges, etc.) to sophisticated statistical and machine learning-based algorithms (including gamut mapping, Bayesian approaches, and neural network-based algorithms). Despite high computational processing power, a solution capable of achieving human-level color constancy independently of lighting and light sensor types has not yet emerged. Currently, there is no method that can perfectly resolve the problem in realistic and natural situations by utilizing both aspects of color adaptation and oppositional process theory. Furthermore, there is no solution that utilizes neural normalization processes based on the responses of adjacent photoreceptors, inspired by the insect eye, to estimate image lighting. For the reasons mentioned above, there is an unmet need in the field of image processing for computationally maintaining human color constancy without the need for lighting adaptation or the use of special light sensors. The present invention provides a novel biomimetic solution to this color constancy challenge and offers stable color appearance and high color discrimination during image processing by applying an algorithm inspired by the visual systems of humans and insects. Furthermore, it is recognized that the success of robot navigation depends on the accuracy of visual position recognition and position estimation. However, the presence of shadows caused by changing lighting conditions poses significant challenges to these processes. Shadows from various light sources cause inconsistencies in color, texture, and shape, leading to inconsistencies in existing visual features across different scenarios. This inconsistency degrades overall navigation efficiency by hindering precise position matching and recognition. In this case, it must be recognized that improving recognition accuracy requires only consistent colors, rather than actual colors. To address these concerns, processing is also required to detect and remove shadows within the context of color constancy to mitigate their negative impact on vision-based navigation systems. Shadows cast by objects and structures under varying lighting angles and intensities significantly alter the appearance of scenes and objects. These variations cause inconsistencies in the visual data captured by the robot's sensors, making it difficult to accurately match features across diverse lighting conditions. Consequently, the reliability and precision of visual position recognition and position estimation are degraded, impairing th